Deep learning-based denoising and computational field-of-view extension towards rapid wide-field mid-infrared...

Recent advancements in wide-field mid-infrared photothermal (MIP) microscopy offer rapid chemical contrast for biological, biomedical, and polymer analysis at submicron resolution. Currently, the field of view (FOV), signal-to-noise ratio (SNR), and imaging speed are limited by mid-infrared pump fluence, non-uniform mid-infrared illumination, relatively small photothermal scattering change, and consequently the need for multiple camera frame averaging to achieve photothermal contrast. To correct non-uniform illumination, expand FOV, and improve SNR, flat-field correction, feature-based image stitching, and deep-learning image denoising are integrated within our previously developed wide-field quantum-cascade-laser MIP workflow. A feature-based stitching algorithm with flat-field correction combines individual small MIP frames into large composite mosaics, which is demonstrated on expanded MIP images of tuberculosis-infected lung tissue. Quantitatively, stitching delivered high-fidelity reconstructions, yielding structural similarity index measure (SSIM) and normalized cross-correlation (NCC) values >0.97 in overlaps. For denoising, a residual dense U-Net (RDUNet) is trained with high-frame-average data to predict high-SNR images from low-frame-average inputs. The method is applied to wide-field MIP images of polyethylene terephthalate (PET) microparticles. RDUNet boosted SSIM by >0.7 and peak SNR by >5 dB while predicting clean images of PET particles in less than 1 second. Overall, stitching of mosaics from multiple wide-field MIP images at low-frame numbers paves the way towards large-area, high-contrast MIP imaging, advancing chemical imaging towards high-throughput applications.